No 145- February 2012

Assessing the Returns to Education in the Gambia

Jeremy D. Foltz and Ousman Gajigo

Correct citation: Foltz, Jeremy D.; and Gajigo, Ousman (2012), Assessing the Returns to Education in the Gambia, Working Paper Series N° 145, African Development Bank, Tunis, Tunisia.

Vencatachellum, Désiré (Chair) Anyanwu, John C. Verdier-Chouchane, Audrey Ngaruko, Floribert Faye, Issa Shimeles, Abebe Salami, Adeleke

Coordinator

Working Papers are available online at

http:/www.afdb.org/

Copyright © 2012

African Development Bank

Angle des l’avenue du Ghana et des rues

Pierre de Coubertin et Hédi Nouira

BP 323 -1002 TUNIS Belvédère (Tunisia)

Tel: +216 71 333 511

Fax: +216 71 351 933

E-mail: afdb@afdb.org

Salami, Adeleke

Editorial Committee Rights and Permissions

All rights reserved.

The text and data in this publication may be

reproduced as long as the source is cited.

Reproduction for commercial purposes is

forbidden.

The Working Paper Series (WPS) is produced

by the Development Research Department

of the African Development Bank. The WPS

disseminates the findings of work in progress,

preliminary research results, and development

experience and lessons, to encourage the

exchange of ideas and innovative thinking

among researchers, development

practitioners, policy makers, and donors. The

findings, interpretations, and conclusions

expressed in the Bank’s WPS are entirely

those of the author(s) and do not necessarily

represent the view of the African Development

Bank, its Board of Directors, or the countries

they represent.

Assessing the Returns to Education in the Gambia

Jeremy D. Foltz and Ousman Gajigo1

1 Jeremy D. Foltz and Ousman Gajigo are respectively Associate Professor, University of Wisconsin, Madison, and

Economist, African Development Bank respectively. The authors are grateful to the Central Bureau of Statistics for

giving access to the household data sets and the Department of State for Education for access to the school data.

They are also indebted to Abu Camara and Lamin Fatty at the Central Bureau of Statistics for the innumerable ways

they helped with the data. This work has also benefited from suggestions by Laura Schechter, Thi Minh Ngo,

seminar participants at Wisconsin and Yale, participants at the 4th Minnesota International Economic Development

Conference, the editor and two anonymous reviewers.

AFRICAN DEVELOPMENT BANK GROUP

Working Paper No. 145 February 2012

Office of the Chief Economist

4

Abstract

Using three nationally representative surveys from the country, we estimate the private

rates of returns to education in The Gambia. To obtain consistent estimates, we exploit

exogenous variation in school availability in the country at the district level at the time

current wage earners where born. Our results show that the private rates of returns to

education are quite high, although heterogeneous across regions of the country. The

high rates of returns are robust to alternate formulations.

Keywords: Gambia; Schooling; Returns to Education; Wage

JEL Codes: C36, I21, I25, J24,

5

5

I. Introduction

The importance of education in development is a perennial topic in economics

especially in the context of sub-Saharan Africa’s development experience. The

connection is not surprising since the region stands out both in its low level of schooling

and its low historic average rate of economic growth. In the macroeconomic growth

literature, Krueger and Lindahl (2001) showed that education is positively associated

with economic growth, a result that accords well with many previous studies. Micro-level

research on private rates of returns to education has shown disparate estimates in sub-

Saharan Africa in the private benefits to education. Our work focuses on private returns

to education in The Gambia2, a small country in West Africa with very low levels of

schooling. Like other countries in the region, it also has achieved little economic growth

since independence in 1965. It is therefore not surprising that the country is not on

schedule to achieve one of the Millennium Development Goals: universal primary

education by 2015.

This work adds to the large literature that provides a range of estimates on the

private rate of returns to education in Africa. Psacharopoulos’ 1994 review of the

literature found that the average private rate of returns to education for sub-Saharan

Africa were around 13%, though with significant variation in estimates across countries.

Many studies on returns to education in sub-Saharan Africa have improved on the

estimation methodology of the works cited by Psacharopoulos (1994). For example,

Glewwe (1996) was able to directly address school quality and ability, problems that

2 The Gambia’s official name is “The Gambia” including the capitalized article, so we use that each time

that way in the text.

6

6

have largely been unaddressed in the earlier literature. That work estimated the rate of

returns to education in Ghana to be in the range of 3% to 6%. In a more recent and

equally rigorous work, Oyelere (2010) estimates returns to schooling that are slightly

under 4% in Nigeria. And in another study about a West African country, Kazianga

(2004) also estimated returns to education in Burkina Faso, finding that the returns to

education are between 9% and 17% at the primary school level and 13% and 20% at the

secondary school level in the private sector. In the public sector, the returns range

between 0% to 6% at the primary level and 10% and 11% at the secondary level. Schultz

(2004) found estimates for Ivory Coast ranging from 3.8% to 28%. In a different region

of sub-Saharan Africa, Siphambe (2000) estimates the rate of returns to education in

Botswana to be in the range of 12% to 18%.

The literature therefore provides a wide range of estimates of the rate of return to

education in sub-Saharan Africa. It could be the case that there are indeed very large

differences between countries in the rates of returns to education since there has been

very little replication of estimates within a single country. Part of the difference in

estimates may also be due to the use of improved econometric techniques among recent

papers. Some of these new approaches have addressed issues such as ability bias and

selection - problems that were not always addressed in many earlier papers.

Another possibility is that differences in estimation strategies can also produce

different results since the estimates may be specific to only a subset of the population in a

given country. Specifically, the estimates from using an instrumental variable approach

may not be comparable across different studies that employ different instruments since

such an estimation strategy produces the local average treatment effects (Card, 2001;

7

7

Imbens and Angrist, 1994). Typical estimates using instrumental variables, in which the

most common instrument measures access to schooling, provide measures of the returns

to schooling for those who would have continued in school but did not have access to

schooling. Given that in the African context there is great variation across countries,

ethnic groups, and religions in the proclivity of parents to send their children to school

even when it is available and affordable, one should also expect great variation in

estimates of returns based on that population.

This work contributes to the literature by providing the first estimates of the

private rate of returns to education for The Gambia and among its regions. Our estimates

rely on the exploitation of the exogenous variations in the availability of schools across

the country at the district level and its interaction with year of birth of individuals to

control for ability bias. In addition, we use exogenous rainfall shocks to control for

selection bias. Like many instrumental variables, ours are not perfect. We discuss the

possible violations of the exclusion restriction and provide further robustness checks to

mitigate against them.

Our study uses three nationally representative household surveys from 1992, 1998

and 2003 that provide a very high coverage rate for the overall population of The

Gambia. The results show high and significant private rate of returns to education for

individuals in the wage sector. The results also suggest large significant differences in the

rate of returns to education across regions.

This work proceeds as follows. In section II, we describe the data set, which

includes the historical education data of The Gambia. Section III discusses our estimation

8

8

strategy. Our main results on private rate of returns are presented in Section IV, where we

also present our robustness checks. Section V concludes the paper.

II. Data description

Like other countries in the Sub-Saharan Africa region, The Gambia is very poor

by world standards, predominantly rural, and has an agricultural-based economy. Since

independence in 1965, economic growth has been nearly non-existent, averaging about

0.7% per year between 1965 and 2009. The GDP per capita (PPP) in 2009 was $1,285 in

constant 2005 US dollars (World Bank 2010).

The data set we use includes three household surveys (1992, 1998 and 2003)

carried out by the Central Bureau of Statistics in The Gambia. These surveys, the

Household Poverty Survey, cover all the seven regional administrative areas (regions 2-6

are commonly known as Divisions) and most districts3. The surveys are repeated cross-

section and are carried out approximately every five years. The numbers of households

sampled in years 1992, 1998 and 2003 were 1,387; 1,923 and 4,672 respectively, making

a pooled sample size of 7,982 households. This household coverage results in 62,538

sampled individuals. Out of this sample, approximately 13,780 individual are wage

earners.

Table 1 presents the summary statistics of key variables for individuals in the

labor market. The three time periods are very similar in most of the variables listed. As

expected, the average years of schooling (S), at 2.92 years, is very low in the sample.

Surprisingly, the average of this variable is lower in the 1998 and 2003 samples than in

3 The country is divided in to roughly six administrative areas: five divisions plus the capital and its

surrounding area called the Greater Banjul Area. Within the five divisions are districts numbering close to

40. The 1992 and 1998 surveys covered most but not all districts. The 2003 survey covered all districts.

9

9

1992, but virtually all of that difference can be accounted for in the differences between

the rural samples of the two sets of periods. While wage earners are similar to the general

sample in average schooling, they are on average 7 years older relative to the general

population. Similarly, women have significantly fewer years of schooling than men,

attaining 2.34 years of schooling on average relative to men’s average of 3.68 years.

In constant 2003 values, annual wages have increased by 49% between 1992 and

2003. Figures 1 and 2 provide the wage distributions by gender and location. Unlike

wages, total household income has not grown over this time period. In fact, it fell by 14%

over that time period most likely because of the drought of 2003. This is consistent with

macroeconomic figures since the average annual GDP per capita growth rate was -0.05%

from 1992 to 2003 (World Bank 2010).

Figure 1: Kernel Densities of Male and Female wages.

10

10

Figure 2: Kernel Densities of Urban and Rural wages.

11

11

Table 1: Summary statistics of some key variables (standard deviations are in parentheses). This

summary is restricted to wage workers we use in our analysis.

Pooled 1992 1998 2003

1 2 3 4 5 6 7 8

Obs.

Mean

(Standard

Deviation)

Obs.

Mean

(Standard

Deviation)

Obs.

Mean

(Standard

Deviation)

Obs.

Mean

(Standard

Deviation)

Log-wage‡ 13,779

8.48

(1.66) 3,048

7.82

(1.67) 3,013

8.29

(1.72) 7,718

8.82

(1.53)

Schooling (S) 13,507 2.92

(4.14) 3,036

3.63

(3.70) 2,956

2.52

(4.37) 7,515

2.87

(4.15)

Age 13,775 29.39

(16.50) 3,048

36.81

(13.26) 3,009

38.65

(14.18) 7,718

27.01

(16.45)

Experience4 (E) 13,775

19.28

(13.38) 3,048

18.91

(13.12) 3,009

20.78

(13.96) 7,718

18.84

(13.21)

Female 13,775 0.47

(0.50) 3,048

0.43

(0.50) 3,009

0.39

(0.49) 7,718

0.48

(0.50)

Rural 13,779 0.56

(0.50) 3,048

0.53

(0.50) 3,013

0.55

(0.50) 7,718

0.57

(0.50)

No Schooling 13,592 0.60

(0.49) 3,048

0.79

(0.42) 2,983

0.73

(0.45) 7,561

0.56

(0.45)

Primary

School 13,592

0.18

(0.38) 3,048

0.05

(0.21) 2,983

0.08

(0.27) 7,561

0.21

(0.41)

Secondary

Schooling 13,592

0.20

(0.40) 3,048

0.13

(0.33) 2,983

0.17

(0.38) 7,561

0.22

(0.41)

Tertiary

Schooling 13,592

0.02

(0.12) 3,048

0.03

(0.16) 2,983

0.02

(0.15) 7,561

0.01

(0.11) ‡Log-wage is the natural log of annual wage in 2003 Dalasis ($1=27 Dalasis in 2003).

No Schooling=1 if no schooling; Primary School=equals 1 if the highest level of schooling is

the primary level; Secondary School=equals 1 if the highest level of school is the secondary

level; and Tertiary Education=equals 1 if the individual has tertiary level education. The

variables No Schooling, Primary School, Secondary School and Tertiary Education sum to 1.

School and Education Data

Our analysis relies heavily on the historical education and population data. The data on

the dates and location of school constructions comes from the Ministry of Education, which

keeps a record of all formal schools ever constructed in the Gambia. For example, the first

4 Experience is measured in years. We do not have data on real labor market experience acquired. Rather, we

constructed it as follows: E=Max{0, age-18}. The choice of 18 is admittedly arbitrary. It is possible to acquire labor

market experience before reaching 18, especially if one never attended school. We have used age instead of this

constructed experience variable and the results are highly similar.

12

12

modern school was constructed in The Gambia in 1835. This is far earlier than the earliest date

of birth for a worker in our sample, which is 1912. Our time series of district and national

population data comes from the Central Bureau of Statistics. Because of its small size (the land

size is 4000 square miles), collecting population data for the Gambia even in early 20th

century

was far less challenging than in many other African countries. Population figures for the-then

Gambia colony5 had been kept by the colonial authorities as early as 1900. In 1900, the

population of The Gambia was approximately 110,000, and at independence in 1965 it had

reached 407,800 people.

Given the history, schooling is naturally low in levels and even in growth.6 The adult

literacy rate in 1990 was 26%. Net enrollment rate at the primary level in 1991 was 48%. The

pupil to teacher ratio increased from 31.3 in 1990 to 32.9 in 1998, due to a slow growth in

enrollment combined with little investment in education, schools, or addition of teachers over

time (World Bank 2010). While historical school attendance and achievement numbers are

difficult to find, data on the dates of construction of every school in the formal education sector

in the country does exist. The levels and densities of both primary and secondary schools from

1900 to 2005 are presented in figures 3 and 4.

In addition to the low total number of schools in the country, the distribution of schools

has also been very unequal across regions, with a bias toward the capital and other urban areas.

Table 2 shows the spatial school density per region. The farther away a region is from the

capital, the smaller the number of schools per square kilometer.

5 The Gambia was a British colony and protectorate. The small territory was divided into two parts: the colony and

the protectorate. The colony covered the capital and coastal areas and was ruled directly by colonial administrators

headed by a governor appointed from London. This area corresponds to the administrative areas of Region 1 and

parts of Region 2. The interior was considered the protectorate, which was ruled indirectly through local chiefs and

corresponds to parts of Region 2, and all of Regions 3, 4, 5 and 6 (Hughes and Perfect 2006). 6 Good data on enrollment, literacy and educational expenditure does not extend earlier than the 1990s.

13

13

Figure 3: The number of primary and secondary schools in the country from 1900-2005.

Source: school data comes from the Ministry of Education while population data comes from the

Central Bureau of Statistics.

Figure 4: The number of schools per 1000 people from 1900 to 2005

Source: School data comes from the Ministry of Education while population data comes from

the Central Bureau of Statistics.

0

20

40

60

80

100

120

140

0

50

100

150

200

250

300

350

400

19

00

19

04

19

08

19

12

19

16

19

20

19

24

19

28

19

32

19

36

19

40

19

44

19

48

19

52

19

56

19

60

19

64

19

68

19

72

19

76

19

80

19

84

19

88

19

92

19

96

20

00

20

04

Seco

nd

ary

Sch

oo

ls

Pri

mar

y Sc

ho

ols

Primary schools

Secondary Schools

0

0.01

0.02

0.03

0.04

0.05

0.06

0.07

0.08

0.09

0

0.05

0.1

0.15

0.2

0.25

0.3

19

00

19

04

19

08

19

12

19

16

19

20

19

24

19

28

19

32

19

36

19

40

19

44

19

48

19

52

19

56

19

60

19

64

19

68

19

72

19

76

19

80

19

84

19

88

19

92

19

96

20

00

20

04

Nu

mb

er

of

Seco

nd

ary

Sch

oo

ls p

er

10

00

P

eo

ple

Nu

mb

er

of

Pri

mar

y Sc

ho

ols

pe

r 1

00

0

Pe

oo

ple

Primary School Density

Secondary School Density

14

14

Table 2: Number of Schools per 10 square kilometers

1965-1969 1970-1974 1975-1979 1980-1984 1985-1989 1990-1994

Primary

Schools

Sec.

Schools

Primary

Schools

Sec.

Schools

Primary

Schools

Sec.

Schools

Primary

Schools

Sec.

Schools

Primary

Schools

Sec.

Schools

Primary

Schools

Sec.

Schools

Region 1 2.258 1.828 2.344 1.914 2.882 1.935 3.355 1.935 3.656 1.978 4.108 2.516

Region 2 0.114 0.012 0.123 0.029 0.136 0.029 0.189 0.029 0.205 0.029 0.215 0.043

Region 3 0.116 0.000 0.123 0.018 0.139 0.019 0.239 0.019 0.271 0.019 0.334 0.023

Region 4 0.049 0.000 0.051 0.005 0.067 0.006 0.172 0.006 0.174 0.009 0.175 0.013

Region 5 0.033 0.000 0.033 0.002 0.046 0.003 0.093 0.003 0.105 0.003 0.109 0.003

Region 6 0.021 0.000 0.026 0.000 0.057 0.000 0.125 0.002 0.155 0.005 0.169 0.005

III. Estimation Strategy and Results

To estimate the private rate of returns to education in the country, we start with the

following standard Mincer-type equation:

2 '

1 2ln( )ikik ik ik ik ikY S E E

, (1)

where ln( )ikY denotes the natural log of wage income, Sik stands for years of schooling attained

and Eik denotes labor market experience for individual i in district k. The vector χ represents

other determinants of earnings such as sex, rural residence, and regional location. Also included

are the two year dummies (1998 and 2003), with 1992 being the excluded year.

A direct estimation of equation (1) would likely lead to a biased and inconsistent estimate

of the rate of returns to education for a number of reasons. The first reason is that individuals in

the wage labor market are unlikely to be a representative sample of the population. Specifically,

participation in the labor market occurs only if the market wage is equal to or exceeds the

individual’s reservation wage. Secondly, since unobserved ability of individuals is likely to be

correlated with schooling and wages, any estimate of from equation (1) would likely suffer

from omitted variable bias.

In order to solve the first potential bias with our data, let the labor market participation

decision of individual i be given by:

15

15

' '

ik ik ik ikL R (2)

where 1ikL if an individual is in the labor market and 0 otherwise, while ikR is a vector of

variables that affect an individual’s decision to enter the labor market through their effect on the

reservation wage but do not directly affect the market wage. In this formulation, 1ikL when we

observe labor market participation indicating that the market wage exceeds the individual’s

reservation wage. In addressing the selection problem, the challenge is to identify variables in R,

which we do by taking advantage of rainfall risk in the area since agriculture is the primary

economic activity. Kijima, Matsumoto and Yamano (2006) in Uganda, Rose (2000) in India,

Cameron and Worswick (2003) in Indonesia and Ito and Kurosaki (2006) in India all show

significant responses of labor supply to rainfall risk in agriculturally dominated areas.

We use four different rainfall variables in R: rainfall shock (in district) in survey year,

rainfall shock the year before, rainfall shock two years before and coefficient of variation of

rainfall (summary statistics of these variables are provided in Table 3). We define a rainfall

shock at the district level as the deviation of the year’s rainfall from the five-year average. The

coefficient of variation is the ratio of the standard deviation to the mean of rainfall in district

over the previous five years. The rainfall shock variables provide aggregate risk to agricultural

production (at the district level) that is unlikely to be mitigated by risk sharing among

households because they face spatially covariant rainfall distributions. While the rainfall shock

variables provide transient risk, the coefficient of variation of rainfall provides a relatively more

permanent measure of risk (Rose 2000). Overall the rainfall variables are likely to affect the

reservation wages of individuals through their effect on agricultural income, but unlikely to have

16

16

a significant effect on current wages. Therefore, their effect is a “push” effect – that is, on labor

supply.

The direction of the effect of rainfall shocks and coefficient of variation on labor market

participation is ambiguous because households in The Gambia, as in many African countries

(Barrett, Reardon and Webb 2001), have diversified livelihood strategies. For example, a

negative rainfall shock can increase labor supply if the prevailing market wage exceeds the

marginal returns to labor on the farm or in the non-farm enterprise operated by the household or

individual. Conversely, a positive rainfall shock can reduce labor supply if the returns to farm

labor exceed the prevailing market wage. Given that most households and individuals do not

exclusively farm (or work exclusively as entrepreneurs or wage laborers), the rainfall effect on

labor supply is an empirical question. The validity of the above exclusion variables (rainfall

shocks and coefficient of variation) depends critically on their lack of direct effects on wages. In

other words, these variables need to have significant effects on labor market participation but no

direct effect on wages. In the appendix (table A1), we show that all the rainfall shock variables

and the coefficient of variation of rainfall have no direct statistically significant effect on log

wages.

There is also a potential “pull” effect – that is, a variable affecting labor demand, which

can determine labor market participation and therefore would belong in R. We therefore add the

proportion who are self-employed in each district as a measure of labor demand. The appendix,

table A1, shows that like the rainfall data, business ownership in the district does not have a

statistically significant effect on wages.

17

17

Using these “push” and “pull” variables, we estimate equation (2) as a Probit model to

obtain the Inverse Mills Ratio from labor market participation. The Inverse Mills Ratio can be

used to augment equation (1) to address the potential selection bias from only observing wages

from the labor market participation. The result in table 4 shows the relevance of the above

selection instruments. A rainfall shock in the survey year has a positive effect7 on labor market

participation but it is not statistically significant, while rainfall shocks in the preceding year and

two years before both have significant and negative effects on labor market participation. We

also found that the coefficient of variation of rainfall decreases labor market participation, which

is consistent with Ito and Kurosaki (2006). And finally, higher numbers of businesses in the

district is correlated with higher labor force participation. Table 4 shows that years of schooling

(S) has a U-shaped relationship with labor force participation. Without adding a quadratic term

for the number of years of schooling as we do in column 2, the result in column 1 alone would

have implied a counter intuitive relationship.

Table 3: Summary Statistics of identifying variables in equation in R in equation (2).

7 In other words, an above average rainfall in district is associated with employment in the wage sector.

Variable Mean

Standard

Deviation

1 2

Coefficient of Variation of Rainfall 0.24 0.13

Current Year Rainfall Shock -16.90 240.06

Rainfall Shock in Preceding year -

166.16 146.61

Rainfall Shock in 2 Years Earlier -41.90 126.34

Proportion of Business Owners in district 0.06 0.13

18

18

Table 4: Relevance of Selection Instruments. Below are the probit results (marginal effects). The

dependent variable is whether the individual is in the labor market. Robust and clustered

standard errors are in parentheses. Note that the number of observations here far exceeds the

sample in the wage regression because we use the whole sample.

Whole Sample

1 2

Schooling (S) -0.002***

(0.0005)

-0.020***

(0.002)

Schooling Squared (S2)

0.002***

(0.0001)

Age 0.041***

(0.0004)

0.037***

(0.001)

Age sq. -0.0004***

(0.0000)

-0.0004***

(0.00001)

Female Dummy -0.109***

(0.004)

-0.112***

(0.004)

Rural Dummy -0.027***

(0.005)

-0.039***

(0.006)

1998 Dummy -0.026*

(0.015)

-0.021*

(0.012)

2003 Dummy -0.027*

(0.016)

-0.017*

(0.009)

Coefficient of Variation of Rainfall -0.152***

(0.032)

-0.581***

(0.091)

Current Year Rainfall Shock 0.00001

(0.0000)

0.0001***

(0.00002)

Rainfall Shock in Preceding year -0.0001**

(0.00003)

-0.0004***

(0.00005)

Rainfall Shock in 2 Years Earlier -0.0002***

(0.00003)

-0.0002***

(0.0001)

Proportion of Business Owners in district 0.292***

(0.046)

0.307***

(0.089)

Observations 50633 50633

Log Likelihood -20253 -20334

***significant at 1% level; **significant at 5% level; *significant at 10% level. The excluded

year dummy is 1992.

19

19

In order to address the possibility of an omitted variable bias due to unobserved ability,

we use an instrumental variable estimation method that exploits the variation in access to

schooling at the time of an individual’s birth. We accomplish this by interacting the primary and

secondary school densities in district with year of birth of each individual8. Our first stage

equation is

2 2 '

1 2 3 4 5ik ik ik ik ik ik ikS P P M M (3)

where ikP and ikM respectively denote the densities of primary and secondary schools in district k

the year individual i was born9. As is evident from equations (1) and (3), the excluded

instruments are the densities of primary (P) and secondary schools (M) and their quadratic terms.

For these variables (P and M) to serve as proper instruments, they most be highly correlated with

S and be uncorrelated with ability in in equation (1)10

.

The first requirement of our identification strategy concerns the relevance of the

instruments. The proximity to schools, which partially proxies the cost of schooling,11

is likely to

directly influence the probability of parents sending their children to school. This requirement is

also directly testable. First, we show that the school proximity is indeed a significant determinant

of schooling as shown in the results in Table 7. Both measures of school density and their

quadratic terms show significant effects on educational attainment among wage earners.

8 In this way, the instruments capture both the effects of school access as well as any year effects. They are similar

in spirit to the instruments used in Duflo (2001) and Oyelere (2010) 9 Our findings in the results section are unchanged if we instead use school densities in districts at the time

individual was 6 years old - that is, a year before students can be enrolled in primary school. 10

This second stage equation (equation 1) does not control for age. This is because age and experience are highly

correlated in our sample (the correlation coefficient is 0.9). However, in a later section, we estimate the wage

equation for different age cohorts. 11

In areas without schools, parents who want to send their children to school will foster them out to other families,

sometimes related, sometimes not, in other towns where there are schools. This child fostering inevitably imposes

additional costs relative to keeping a child at home and sending them to a nearby school.

20

20

Furthermore, in the instrumental variable estimation of equations (1) and (3) jointly, we provide

the F-test results for the joint statistical significance of 2 3 4, , and 5 . These coefficients are

jointly significant statistically, showing that our instruments are highly correlated with the level

of schooling attained.

Another condition for the consistent estimation of is that these instruments (P and M)

are uncorrelated with ability, which is relegated in ik . In other words, this exogeneity

assumption implies that school density in the district where the individuals were born is

uncorrelated with ability and current wages. While this requirement is not directly testable, we

make the case that the likelihood of the condition being violated is low in our empirical strategy.

Because school density differs from district to district and is systematic while ability is likely to

be randomly distributed in the population, it is unlikely that variations in the density of schools

are correlated with an unobservable such as ability.12

As table 2 makes clear, the number and

density of schools have been very low in The Gambia. Because the exposure to schooling is very

limited in all regions of the country, the low average level of schooling makes it unlikely that

only high ability individuals would have access to and attend school.

On the requirement that there should be no correlation between school density in an

individual’s district and her current wage, we acknowledge that this probability is not as low as

that of the requirement that ability and school density having low correlation. We, nevertheless,

present some evidence that observed current wages are not likely to be correlated with school

density in individual’s district at her time of birth. While the correlation of school densities over

12

It is unlikely that the colonial government was responsive to local needs in terms of where schools should be built.

And while it is likely that the post-independence governments will favor urban areas for many public projects, it is

hard to see how the distribution of these public projects is correlated with ability.

21

21

time is unavoidable, we show in table 5 that there is a convergence in density of schools per

region. For example, while Region 3 had the second lowest primary school density in 1960, it

ended up having the highest density in 1990. In other words, while regions close to the coast had

a relatively higher number of schools initially, they have also experienced tremendous increases

in population as shown by their explosion in population density between 1960 and 1990. And

during this same period, the number of schools in the more remote regions has started to increase

significantly. So while we acknowledge the possibility of correlation between school density in

district at the time of the individual’s birth and her current wage, any such correlation is likely to

be mitigated by rapid population growth.

Table 5: Primary School and Population Density in Regions between 1960 and 1990.

School Density (per 1000 people) Population Density (per sq. km)

1 2 3 4 5 6 7 8

1960 1970 1980 1990 1960 1970 1980 1990

Region1 Primary 0.281 0.134 0.112 0.090

323

840

1427

2906 Secondary 0.211 0.104 0.064 0.051

Region2 Primary 0.236 0.299 0.267 0.209

29

53

70

138 Secondary 0 0.075 0.049 0.039

Region3 Primary 0.149 0.225 0.272 0.307

24

27

47

42 Secondary 0 0.024 0.028 0.021

Region4 Primary 0.126 0.250 0.269 0.270

20

25

32

101 Secondary 0 0 0.019 0.032

Region5 Primary 0.201 0.143 0.202 0.258

21

33

41

51 Secondary 0.017 0.022 0.025 0.020

Region6 Primary 0.054 0.064 0.154 0.218

27

43

50

78 Secondary 0 0 0 0.007

22

22

Table 6: Summary Statistics of Excluded Instruments

Variable Obs. Mean

Standard

Deviation

1 2 3

Primary School

Density (P) 13770 0.222 0.124

Secondary School

Density (M) 13728 0.092 0.117

Primary School Density (P) is the number of primary schools per 1000 people in district when

individual was born.

Secondary School Density (M) is the number of secondary schools per 1000 people in district

when individual was born.

Even after estimating returns to education with the above corrections and adjustments, we

are still left with the issue of how to properly interpret the estimate for in equation (1) after an

IV estimation. The consistent estimation of requires that the marginal returns to education be

similar for all individuals (Card 2001). However, if returns to education are heterogeneous across

individuals, then the results of our IV estimation would give us the weighted average of the

returns to schooling of individuals induced to attend school by the increase in school densities.

This latter outcome would be equivalent to local average treatment effect or LATE (Card 2001;

Deaton 2010; Imbens and Angrist 1994). In other words, our estimates of the returns to

education may be restricted to the sub-sample of the population on the cusp of school attendance

or enrollment and would be driven by the elasticity of that enrollment with respect to access.

This latter point has some implication for the comparability of estimates of returns to education

across different studies using different estimation strategies. As long as different instrumental

variables (in this paper it is school construction but could be lower school fees in others) have

different elasticities, the comparability of estimated returns to education would be limited. In

23

23

other words, the LATE will likely differ based on the type of instrument used and its effect on

the level or years of education attained. This should be kept in mind in comparing returns to

education across countries and studies.

Table 7: Determinants of Schooling Attainment (for wage earners only). The dependent variable

is the years of schooling (S). Robust and clustered standard errors are in parentheses. See

previous the preceding figures in table 6 for summary statistics of variables P and M.

OLS

1

Age 0.013

(0.010)

Age squared -0.001

(0.0001)***

Rural Dummy -1.252

(0.095)***

Female Dummy -1.535

(0.067)***

1998 Dummy -1.012

(0.092)***

2003 Dummy -1.289

(0.079)***

Secondary School

Density (M)

14.364

(2.085)***

Secondary School

Density Sq. (M2)

-20.136

(4.681)***

Primary School

Density (P)

3.069

(1.456)**

Primary School

Density (P2)

-6.546

(3.639)*

Constant 4.445

(0.395)***

Regional dummies Yes

Observations 13457

F-test p-values§

23.11

(0.000)

R squared 0.19

***significant at 1% level; **significant at 5% level; *significant at 10% level. The excluded

year dummy is 1992. §The F-test is the test of the joint significance of the coefficients on M, M

2, P and P

2.

24

24

IV Results:

Table 8 presents our estimates of the private rate of returns to education in the wage

sector in The Gambia. This table shows our second stage results from equation (1) and also the

first stage results (equation 3). The p-values of the Hansen-Sargan statistics validate our

excluded instruments and the high value of the F-test statistic of the joint significance suggests

they are highly relevant. The statistical significance of the Inverse Mills Ratio also suggests that

the sample selection effect is non-trivial.

The OLS estimate of the rate of returns to schooling for the pooled sample is 6.8%

without selection correction and 7.4% with selection correction, which are not significantly

different from each other statistically. The IV estimates of the rate of returns to schooling are

24.1%, 19.1%, and 35.8% depending on specification. Our preferred rate of returns to education

for the country is 24.1% in column 3 of table 8 because this particular result addresses both

selection and omitted variable bias.

The results by gender, in table 8, show that the rate of returns to education is higher for

males (35.8%) than for females (16%)13

. This is surprising because the average level of

schooling of males is higher than that of females in the sample. Assuming diminishing marginal

returns, then one would expect the returns to be higher for females relative to males. However,

the expectation of higher female marginal returns assumes the absence of unobserved gender

discrimination. Specifically, societal norms could counteract the expected higher female returns

by creating gender barriers in certain higher-return occupations. In an experimental study of

13

We use interaction terms in columns 4 and 5 of table 8 to allow the marginal returns to education to vary by

gender and location. While this procedure forces the coefficients on the other variables to be the same for both

gender and rural dummy, this decision is not very limiting since we are not particularly interested in gender

differences in variables such as experience, rural and year dummies.

25

25

returns to capital among Sri Lankan micro-entrepreneurs, Del Mel et al. (2009) found that the

marginal returns to capital is higher for males than females even though the former have larger

capital sizes. While they found that this gender difference in marginal returns cannot be

explained by differences in ability or risk preferences, it is consistent with differential

concentration in industries and relative intra-household bargaining power. It is also worth

pointing out that Schultz (2004) also found that the returns to education for males are higher than

females across age cohorts in another country in West Africa, Ivory Coast.

The higher marginal returns to education for urban residents (19.1%) relative to rural

workers (7.3%) is also counter-intuitive given the relatively higher average educational

attainment in urban areas. While we would expect higher marginal rate of returns at lower levels

of schooling, the kinds of occupations available in rural areas are unlikely to provide high returns

to education relative to those in urban areas. For example, high return government jobs are more

likely to be located in urban areas. The significant difference in returns to education between

rural and urban areas suggests some heterogeneity across regions. We present the estimations by

regions in table 9 and indeed find large heterogeneity. It appears that the high returns to

education are driven primarily by 2 out of the 6 regions14

. For the four other regions, the rate of

returns to education ranges from 8% to 16%.

The heterogeneity of the rate of returns within a small country such as the Gambia

suggests that the variation in returns within individual countries may be higher than variability

between countries (Psacharopoulos, 1994). Our estimate of the rate of returns to education in

14

The population share of regions 1, 2, 3, 4, 5 and 6 are 26%, 29%, 13%, 5%, 14% and 13% respectively.

26

26

four out of six regions is very similar to the single digit estimated values for other West African

countries in the region (see for example: Oyelere, 2010; Kazianga, 2004 and Glewwe, 1996).

What accounts for the heterogeneity in our estimated rate of returns to education when

we use the IV approach? As can been seen in table 9, our measure of the rate of returns to

education with OLS is very similar across regions. The interpretation of our estimated returns to

education is affected by the potential heterogeneity of the marginal returns to schooling induced

by possibly different responses to school construction and access. Given that different regions of

The Gambia had different levels of school densities at any point in time, it is unlikely that

changes in school densities would induce identical enrollment in all districts and regions. Given

this fact, the appropriate interpretation of our IV estimate is the local average treatment effect

(Card 2001; Deaton 2009; Imbens and Angrist 1994). In other words, our estimate provides the

returns to education on the subset of the population who attained some schooling but would

otherwise not have, had there been no change in construction of schools in their districts. This

interpretation of the IV estimate also suggests that there may be limits to comparing estimated

returns to education across different countries due to differences in estimation strategies. Even

studies carried out on the same population that use different instruments are unlikely to have

comparable estimates of the return to an additional year of schooling.

Another factor that can account for the significant differences in returns to education

across regions is the violation of the exclusion restriction in our IV estimate strategy.

Specifically, historical school densities may affect current wages through channels not restricted

to attained schooling (S). This could result from differences in levels of regional development or

27

27

urbanization - initial differences that can persist through time across regions and districts15

. This

is consistent with the low Hansen-Sargan statistics in column 4 of table 8 and the results for

regions 4 and 5 in table 9. However, even when we exclude regions 4 and 5, our returns to

education estimate for the remaining sample is 24%, almost identical with the value in column 3

of table 8.

15

We would like to thank an anonymous reviewer for alerting us to this possibility and its implications for our

results.

28

28

Table 8: Wage regression results. The dependent variable is the natural log of wage income. In

obtaining returns to educations by gender and location (columns 4 and 5), note the interaction

terms. Robust standard errors (clustered at district level) are in parentheses.

OLS IV

1 2 3 4 5

Schooling (S) 0.068***

(0.005)

0.074***

(0.004)

0.241***

(0.064)

0.191**

(0.074)

0.358***

(0.108)

Experience (E) 0.049***

(0.005)

0.174***

(0.053)

0.366***

(0.117)

0.297**

(0.119)

0.333***

(0.105)

Experience Sq. -0.001***

(0.000)

-0.003***

(0.001)

-0.006***

(0.002)

-0.005**

(0.002)

-0.005***

(0.002)

Female Dummy -0.737***

(0.088)

-1.078***

(0.126)

-1.370***

(0.245)

-1.349***

(0.266)

-0.733***

(0.179)

Rural Dummy -0.603**

(0.230)

-0.580**

(0.227)

-0.339

(0.258)

-0.263

(0.363)

-0.278

(0.288)

1998 Dummy 0.421***

(0.085)

0.193

(0.139)

-0.044

(0.211)

-0.014

(0.223)

0.082

(0.198)

2003 Dummy 0.899***

(0.138)

0.670***

(0.126)

0.466***

(0.129)

0.511***

(0.160)

0.596***

(0.123)

Schooling*Rural -0.118

(0.077)

Schooling*Female -0.190**

(0.083)

Inverse Mills Ratio No 1.368***

(0.573)

3.460**

(1.307)

2.726**

(1.327)

3.093**

(1.175)

Constant 8.204***

(0.073)

5.914***

(0.972)

1.346

(2.526)

3.001

(2.521)

1.362

(2.467)

First Stage

Primary School

Density (P)

3.094

(1.563)**

3.158

(1.611)**

3.156

(1.586)**

Primary School

Density sq. (P2)

-6.530

(3.819)*

-6.546

(3.482)*

-6.539

(3.653)*

Secondary School

density (M)

14.456

(3.623)***

14.461

(2.175)***

14.377

(3.423)***

Secondary School

density (M2)

-20.069

(6.299)***

-20.072

(6.049)***

-20.123

(5.967)***

R squared 0.39 0.38

Regional Dummies Yes Yes Yes Yes Yes

Observations 13503 13033 12989 12989 12989

Hansen-Sargan†

statistic (p-value)

5.59

(0.13)

8.18

(0.04)

4.52

(0.21)

F-test statistic§

(p-value)

23.10

(0.00)

17.08

(0.00)

14.22

(0.00)

***significant at 1% level; **significant at 5% level; *significant at 10% level. The excluded year dummy is 1992. §This is the test on whether the excluded instruments (which appear only in the first stage) are jointly significant.

†This is Hansen-Sargan test with the null hypothesis that the instruments are valid.

29

29

Table 9: The Rate of Returns to Education by Regions. The dependent variable is log of wage income. All controls in Table 8 are also included here.

Region 1 Region2 Region3 Region4 Region5 Region6

OLS IV OLS IV OLS IV OLS IV OLS IV OLS IV

1 2 3 4 5 6 7 8 9 10 11 12

Schooling(S) 0.06*** 0.08** 0.06*** 0.48** 0.06*** 0.14** 0.06*** 0.16* 0.11*** 0.43** 0.07*** 0.14**

First Stage

Primary

School

density (P)

3.42** 3.08*** 5.60** 6.68** 3.24** 7.40**

Primary

School

density sq.

(P2)

-5.61* -6.45* -5.80** -3.887** -5.96* -3.16*

Secondary

School

density (M)

15.34** 14.34*** 16.64* 16.67** 14.58** 17.52**

Secondary

School

density sq.

(M2)

-19.83* -20.19* -18.27* -17.33 -17.97 -17.15**

F-Test

(p-value)

24.02

(0.00)

23.65

(0.00)

25.53

(0.00)

25.53

(0.00)

23.63

(0.00)

21.08

(0.00)

Hansen-

Sargan

statistic (p-

value)

4.27

(0.23)

3.71

(0.30)

5.70

(0.12)

6.72

(0.08)

9.35

(0.02)

5.48

(0.14)

Uses Controls

in Column 3

of Table 8

Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes

Observations 3664 3658 4004 3985 1626 1622 496 493 2234 2225 1479 1474

R sq. 0.18 0.21 0.31 0.48 0.24 0.35

***significant at 1% level; **significant at 5% level; *significant at 10% level.

30

30

Robustness Checks on Rate of Return to Schooling

Since the effect of differences in learning in first and second grade is not the same as in

sixth and seventh grade, it is also possible that the rates of returns to schooling differ at various

levels of schooling. In our main estimation, we did not include higher order terms for the years

of schooling (S) variable in the second stage estimation (equation 1). To determine if this is the

case, we estimate the relationship between years of schooling and log wage with a semi-

parametric technique that does not force any linearity assumption between the two variables.

Equation (4) estimates a semi-parametric equation, which admits a flexible function of

years of schooling, ( )if S and allows other relevant individual variables (age, experience, gender,

rural, regional and year dummies) to enter linearly (Lokshin 2006).

'ln ( )i i i iY f S X (4)

X represents a vector of individual variables and i is a zero-mean error term. The variables that

enter linearly are the same controls in the results presented in column 3 of Table 8 except for the

interaction terms and the inverse mills ratio. The semi-parametric procedure estimates

'ln i i iY X and then smooths 'ln ( )i i i iY X f S (where is the estimation of )

using locally weighted scatter-plot regression (LOWESS).

Figure 5 presents the non-parametric function, ( )if S that results from estimating equation

(4). The estimated figure is approximately linear and does not show dramatic changes in slope. It

is therefore unlikely that our earlier estimation obscures significant heterogeneity in the rate of

returns at different levels of schooling.

31

31

Figure 5: Semi-parametric estimation of the relationship between schooling and log-wage

(N=13,496).

However, the above semi-parametric figure does not address the potential bias that

motivated our IV approach. Following the literature, we also estimate returns to education using

levels of schooling rather than years of education. This could be important if the attainment of

certain educational levels comes with a credential effect. For example, while an individual with 5

years of schooling is only 2 years behind another with 7 years, the differences in the labor market

could be significant given the latter is in possession of a primary school leaving certificate.

To estimate returns to education using levels, S in equations (1) and (3) is now considered

a vector of schooling levels. These levels of schooling, which are dummy variables, are primary

(1 to 6 years of schooling), secondary (7 to 12 years of schooling) and tertiary (greater than 12

years of schooling). The excluded dummy is no schooling (0 years of schooling).The average

05

10

15

log-w

age (

2003 D

ala

sis

)

0 5 10 15 20years of schooling

bandwidth = .4

32

32

years of schooling for those with primary, secondary and tertiary education levels are 3.4, 9.6

and 14.4 respectively.

Table 10 presents the results of our estimation using levels of schooling. As with our

results from using years of schooling, the IV estimates are much higher than the OLS estimates.

However, the OLS estimates in tables 10 are similar to those in table 8 following the method by

Shultz (2004). For example, calculating the implied returns to an additional year of education for

secondary level using the OLS estimates is: (1.173-0.144)/(9.6-3.4)=0.08. Following the same

calculation, it can be shown that the returns to a year of education at the primary and tertiary

levels for the OLS results are 0.05 and 0.11 respectively. For the IV results, the implied estimates

for a year of schooling at the primary, secondary and tertiary levels are 0.24, 0.13 and 0.21

respectively.

33

33

Table 10: Returns to Education by Level of schooling. The dependent variable is log wage. The

reference education level (omitted category) is no school. Robust standard errors (clustered at

district level) are in parentheses.

OLS

IV

1

2

Primary 0.144

(0.072)**

0.825

(0.415)**

Secondary 0.647

(0.052)***

1.607

(0.673)**

Tertiary 1.173

(0.064)***

2.633

(1.471)*

Experience 0.052

(0.015)***

0.164

(0.052)***

Experience Sq. -0.003

(0.001)***

-0.001

(0.0002)***

Female Dummy -0.386

(0.193)**

-1.070

(0.131)***

Rural Dummy -0.600

(0.224)**

-0.290

(0.227)

1998 Dummy 0.202

(0.137)

0.260

(0.173)

2003 Dummy 0.673

(0.129)***

0.793

(0.143)***

Inverse Mills Ratio 1.276

(0.568)***

2.845

(1.221)**

Constant 6.149

(0.961)***

6.815

(0.743)***

Observations 13503

12989

R Squared 0.37

First Stage

Primary

Dummy

Secondary

Dummy

Tertiary

Dummy

Primary School

Density (P)

0.143

(0.046)***

0.211

(0.114)*

0.165

(0.497)

Primary School

Density sq. (P2)

-0.028

(0.014)**

0.111

(0.080)*

0.012

(0.018)

Secondary School

Density (M)

0.131

(0.060)**

0.166

(0.083)**

0.198

(0.094)**

Secondary School

Density (M2)

-0.012

(0.009)

-0.201

(0.106)*

-0.172

(0.090)**

F-test statistic§

(p-value)

21.67

(0.000)

18.75

(0.000)

15.83

(0.000) Hansen Sargan

†

(p-value)

0.27

(0.59)

***significant at 1% level; **significant at 5% level; *significant at 10% level. The excluded year dummy is

1992.§This is the test on whether the excluded instruments (which appear only in the first stage) are jointly

significant.† This is Hansen-Sargan test with the null hypothesis that the instruments are valid.

34

34

There is another potential issue with our estimates of returns to schooling: school quality.

It is possible that differences in school quality may affect the rate at which parents send their

children to school. If there are significant differences in school quality and it is not controlled

for, this could bias our results.

Our data did not include any measure of school quality, such as student teacher ratios,

which are commonly used in the literature, and we therefore cannot directly assess the effect of

school quality. But because the available evidences suggests few year to year differences in

student teacher ratios, we do not suspect that controlling for school quality through student

teacher ratios as commonly practiced in the literature would significantly change our results.

Nevertheless, if there were significant school quality changes over time we hypothesize

that they would be evident in different returns across student cohorts. If school quality changed

significantly over time, it is plausible to suspect that this would be reflected in different rates of

returns to education across different age cohorts. Table 11 shows the results of our main wage

regression run separately for different age cohorts. Returns to education are still significant and

high for all cohorts. Furthermore, none of the estimated coefficients of schooling (i.e. the rate of

returns to education) of the different cohorts are statistically different from our main estimate

(24.1%) in column 3 of table 8.

35

35

Table 11: Returns to Education by Age Cohorts. This regression has the same controls in Table

8 but we only show the coefficient on schooling (S). Robust standard errors (clustered at the

district level) are in parentheses.

IV

Age 25 or

younger

Older than 25 but

younger than or

equal to 35

Older than 35 but

younger than or

equal to 45

Older

than

45 years

1 2 3 4

Schooling (S) 0.232**

(0.069)

0.331***

(0.098)

0.193**

(0.078)

0.241**

(0.098)

Observations 2622 4444 3239 3152

Hansen-Sagan†

statistic

(p-value)

1.04

(0.79)

0.84

(0.83)

1.11

(0.77)

1.22

(0.74)

First Stage

Primary School

Density (P)

3.541

(1.069)***

3.969

(1.195)***

3.019

(1.003)***

3.518

(1.067)***

Primary School

Density sq. (P2)

-4.377

(2.199)**

-3.721

(1.066)***

-5.091

(1.562)***

-4.380

(2.212)**

Secondary School

Density (M)

14.905

(6.714)**

15.315

(7.696)**

14.514

(7.367)**

14.899

(6.681)**

Primary School

Density sq. (M2)

-17.914

(10.178)*

-17.684

(8.976)**

-18.327

(9.118)**

-17.908

(10.005)*

F-test§

Statistic

(p-value)

10.64

(0.00)

7.78

(0.00)

6.19

(0.00)

20.37

(0.00)

Uses Controls in

Column 3 of Table 8?

Yes Yes Yes Yes

***significant at 1% level, **significant at 5% level, *significant at 10% level §This is the test on whether the excluded instruments (which appear only in the first stage) are

jointly significant. †This is Hansen-Sargan test with the null hypothesis that the instruments are valid.

Discussion

All our IV estimates show high rates of returns to education. This leads to the question of

why we observe such a high rate of return and yet low school attendance. The first possible

answer is that sending a child to school remains very costly for the average Gambian household.

The direct cost of schooling involves not only uniforms, books, paper, and other stationery

supplies, but also school fees at the middle and high school levels. The indirect cost stems from

36

36

the loss of child labor on the family farm and as household labor for getting wood and water,

which can be substantial for rural households. To the extent that households are credit

constrained they are also unlikely to be able to borrow against the future increased earnings to

finance current educational expenses.

Another obstacle to schooling is the possibility of perceived low returns to schooling.

This may seem surprising since we just documented high returns to schooling in the wage sector.

However, the wage sector is relatively small in the country. Farming is the more dominant

livelihood in the Gambia. In anticipating future returns to schooling, the agricultural sector and

the informal work sectors are likely to be weighted more than the formal wage sector in a

parents’ expectation of future returns to education.

The preceding paragraph begs the question of what is the reason behind the likely low

returns to schooling in the agricultural or informal sectors. One obvious reason to expect high

returns to education in the agricultural sector is through its link to increased technology adoption.

But there has been no significant technical change in the agricultural sector in The Gambia.

Chavas et al. (2005) found that technical efficiency in Gambian farm households is almost 100%.

This is what one would expect when there is little or no new technology to learn and thus little

heterogeneity in technique or ability to master the technology.

V. Conclusion

We provide the first estimates of the private rate of returns to education in The Gambia.

Using three nationally representative household surveys and exploiting the exogenous variations

in the availability of schools across regions when individuals were born, we are able to obtain

consistent estimates of returns to education. Our IV estimate of the rate of returns to education

for an additional year of schooling is 24.1%. However, this figure seems to be masking some

37

37

significant heterogeneity across regions in the country. In most regions, the rate of returns to

education ranges from 8% to 16%. This figure is higher than other estimates of the private rate of

returns to education in developing countries in general (Psacharopoulos, 1994) and many recent

estimates for West Africa in particular. It is also worth point out that the high inter-regional

differences in returns to schooling could be due to the violation of the exclusion restrictions in

some districts rather than inherent differences in rates of returns to education.

The combination of high estimated returns to education with low levels of school

attendance that are evident in our results suggest that the presence of constraints may prevent

households from fully exploiting the high returns to schooling. School attendance is highly

correlated with proximity to schools and parents directly list cost as one of the reasons for not

enrolling their children. Our results are also consistent with the untested possibility that

households discount the high rate of returns to education in the wage sector because it is a very

small sector relative to agriculture in the Gambian economy. This effect may be exacerbated by

the fact that the agricultural sector in the country has not experienced significant technical

change that is likely to reward education.

The results presented here imply that there is a large scope for interventions in the

education sector to have significant benefits in The Gambia. Most directly, improving access to

schools through construction and staffing of schools as well as reducing direct and indirect costs

of schooling can have direct effects on children’s propensity to attend and have long-term returns

for individuals and the country. In terms of future research, this work raises a number of

questions about the returns to schooling in agriculture and the informal sector. It also raises a

number of important questions on the tradeoffs of schooling and child labor in The Gambia.

38

38

Future work investigating the effects of school access on child labor use in West Africa would

also be a welcome addition to the literature.

References

Barrett, C.B., T. Reardon and P. Webb. 2001. “Nonfarm Income Diversification and Household

Livelihood Strategies in Rural Africa: Concepts, Dynamics and Policy Implications”, Food

Policy, 26(4): 315-331.

Baum, C. F., M.E. Schaffer and S. Stillman. 2007. “Enhanced Routines for Instrumental

Variables/Generalized Methods of Moments Estimation”, Stata Journal, 7(4): 465-506.

Cameron, L.A. and C. Worswick. 2003. “The Labor Market as a Smoothing Device: Labor

Supply Response to Crop Loss” Review of Development Economics, 7(2): 327-341.

Card, D. 2001. “Estimating the Return to Schooling: Progress on Some Persistent Econometric

Problems”, Econometrica, 69(5): 1127-1160.

Chavas, J.P., R. Petrie and M. Roth. 2005. “Farm Household Production Efficiency: Evidence

from The Gambia” American Journal of Agricultural Economics, 87: 160-179.

Deaton, A. 2010. “Instruments, Randomization, and Learning about Development”, Journal of

Economic Literature, 48: 424-455.

Del Mel, S., D. Mckenzie and C. Woodruff. 2009. “Are Women More Credit Constrained?

Experimental Evidence on Gender and Microenterprise Returns”, American Economic Journal:

Applied Economics, 1(3):1-32.

Duflo, E. 2001. “Schooling and Labor Market Consequences of School Construction

in Indonesia: Evidence from an Unusual Policy Experiment” American Economic

Review, 91(4): 795-813.

Glewwe, P. 1996.“The Relevance of Standard Estimates of Rates of Return to

Schooling for Education Policy: A Critical Assessment." Journal of Development

Economics, 51(2): 267-290.

Hughes, D. and D. Perfect. 2006. “A Political History of The Gambia, 1816-1994”, University of

Rochester Press.

Imbens, G. and J. Angrist. 1994. “Identification and Estimation of Local Average Treatment

Effects”, Econometrica, 62(2): 467-476.

Ito, T. and T. Kurosaki. 2006. “Weather Risk and the Off-farm Labor Supply of Agricultural

Households in India”, Hitotsubashi University.

39

39

Kazianga, H. 2004. “Schooling Returns for Wage Earners in Burkina Faso: Evidence from the

1994 and 1998 National Survey ", Economic Growth Center, Yale University.

Kijima, Y., T. Matsumoto and T. Yamano. 2006. “Non-farm Employment, Agricultural Shocks,

and Poverty Dynamics: Evidence from Rural Uganda”, Agricultural Economics, 35(3): 459-467.

Krueger, A. and M. Lindahl. 2001. “Education for Growth: Why and for Whom”, Journal of

Economic Literature, 39:1101-1136.

Lokshin, M. 2006. “Difference-based semiparametric estimation of partial linear regression

models” Stata Journal, 6(3): 377-383.

Mwabu, G. and P. Schultz. 1996. “Education Returns across Quantiles of the Wage Function:

Alternative Explanations for Returns to Education by Race in South Africa”.

American Economic Review Papers and Proceedings, 86(2):335- 339.

Oyelere, R.U. 2010. “Africa’s Education Enigma? The Nigerian Story”, Journal of Development

Economics, 91(1): 128-139.

Rose, E. 2000. “Ex-ante and ex post Labor Supply Response to Risk in a Low-Income Area”,

Journal of Development Economics, 64(2): 371-388.

Psacharopoulos, G. 1994. “Returns to Investment in Education: A Global Update." World

Development, 22(9):1325-43.

Schultz, P. 2004. “Evidence of Returns to Schooling in Africa from Household Surveys:

Monitoring and Restructuring the Market for Education”, Journal of African Economies,

13(2):95-148.

Siphambe, H.K. 2000. “Rates of Return to Education in Botswana”, Economics of Education

Review, 19(3): 291-300.

World Bank, 2010.World Development Indicators, Accessed online at Woldbank.org.

40

40

Table A1: The dependent variable is log wage. The result in this table supports our identification

strategy to control for selection. The excluded variables in the first stage selection estimation,

equation 2 (Biz ownership, shock_t_0, shock_t_1, shock_t_2 and CV_Rainfall) show no direct

significant effect on log wages.

OLS

1 2 3

Current Year Rainfall Shock

(shock_t_0)

0.0001

(0.0004)

0.0001

(0.001)

Rainfall Shock in Preceding year

(shock_t_1)

-0.0003

(0.001)

-0.001

(0.010)

Rainfall Shock 2 Years Earlier

(shock_t_2)

0.001

(0.001)

-0.006

(0.005)

Coefficient of Variation of Rainfall

(CV_Rainfall)

0.632

(0.971)

0.628

(0.966)

Proportion of Business Owners in

district (Biz_Ownership)

0.016

(0.286)

0.021

(0.105)

Schooling (S) 0.066

(0.006)***

0.073

(0.006)***

0.072

(0.008)***

Experience (E) 0.049

(0.005)***

0.044

(0.005)

0.040

(0.004)***

Experience Squared -0.001

(0.0001)***

-0.001

(0.0001)***

-0.001

(0.0001)***

Female Dummy -0.727

(0.087)***

-0.781

(0.085)***

-0.720

(0.097)***

Rural Dummy -0.615

(0.217)**

-0.598

(0.239)***

-0.611

(0.210)**

1998 Dummy 0.543

(0.197)**

0.560

(0.243)***

0.550

(0.204)**

2003 Dummy 0.957

(0.338)**

0.911

(0.314)***

0.950

(0.338)**

Constant 7.936

(0.356)***

7.823

(0.330)***

7.878

(0.318)***

Regional Dummies Yes Yes Yes

Observations 13033 13033 13033

R squared 0.37 0.37 0.37

***significant at 1% level, **significant at 5% level, *significant at 10% level

41

41

Recent Publications in the Series

nº Year Author(s) Title

144 2011 Gabriel Mougani

An Analysis of the Impact of Financial Integration on

Economic Activity and Macroeconomic Volatility in Africa

within the Financial Globalization Context

143 2011 Thouraya Triki and Olfa Maalaoui Chun Does Good Governance Create Value for International Acquirers in Africa: evidence from US acquisitions

142 2011 Thouraya Triki and Issa Faye Africa’s Quest for Development: Can Sovereign Wealth

Funds help?

141 2011 Guy Blaise Nkamleu, Ignacio Tourino and

James Edwin

Always Late: Measures and Determinants of Disbursement

Delays at the African Development Bank

140 2011

Adeleke Salami, Marco Stampini, Abdul

Kamara, Caroline Sullivan and Regassa

Namara

Development Aid and Access to Water and Sanitation in

sub-Saharan Africa

139 2011 Giovanni Caggiano and Pietro Calice The Macroeconomic Impact Of Higher Capital Ratios On African Economies

138 2011 Cédric Achille Mbeng Mezui

Politique Economique Et Facteurs Institutionnels Dans Le

Développement Des Marchés Obligataires Domestiques De

La Zone CFA

137 2011 Thierry Kangoye Does Aid Unpredictability Weaken Governance?

New Evidence from Developing Countries

136 2011 John C. Anyanwu Determinants Of Foreign Direct Investment Inflows To

Africa, 1980-2007

135 2011 John C. Anyanwu International Remittances and Income Inequality in Africa

42

42